aarch64-softfloat: forbid enabling the neon target feature

Author: RalfJung

compiler/rustc_target/src/target_features.rs

@@ -803,14 +803,16 @@ impl Target {
                 // the use of soft-float, so all we can do here is some crude hacks.
                 match &*self.abi {
                     "softfloat" => {
-                        // This is not fully correct, LLVM actually doesn't let us enforce the softfloat
-                        // ABI properly... see <https://github.com/rust-lang/rust/issues/134375>.
-                        // FIXME: should we forbid "neon" here? But that would be a breaking change.
-                        NOTHING
+                        // LLVM will use float registers when `fp-armv8` is available, e.g. for
+                        // calls to built-ins. The only way to ensure a consistent softfloat ABI
+                        // on aarch64 is to never enable `fp-armv8`, so we enforce that.
+                        // In Rust we tie `neon` and `fp-armv8` together, therefore `neon` is the
+                        // feature we have to mark as incompatible.
+                        FeatureConstraints { required: &[], incompatible: &["neon"] }
                     }
                     _ => {
                         // Everything else is assumed to use a hardfloat ABI. neon and fp-armv8 must be enabled.
-                        // These are Rust feature names and we use "neon" to control both of them.
+                        // The return value uses are Rust feature names, hence only "neon" shows up.
                         FeatureConstraints { required: &["neon"], incompatible: &[] }
                     }
                 }